Add Locally disabling gradient computation to PyTorch Frontend about ivy HOT 5 OPEN

Thank you for the detailed answer, I will keep an eye on this and will try to find other issues to work on!

Meanwhile for the purpose of making this issue easier for future contributors, when the functions are ready. The functions in the body could be implemented in this file:
https://github.com/unifyai/ivy/blob/master/ivy/functional/frontends/torch/locally_disabling_gradient_computation.py
and the test could be written here:
https://github.com/unifyai/ivy/blob/master/ivy_tests/test_ivy/test_frontends/test_torch/test_locally_disabling_gradient_computation.py

from ivy.

For this task, I just want to make sure, should we have an experimental function first and then work on it?
I am thinking about this since this functions will change the whole behavior of the library:

from ivy.

For this task, I just want to make sure, should we have an experimental function first and then work on it? I am thinking about this since this functions will change the whole behavior of the library:

Hi @a0m0rajab, thanks for your interest in this task! However, this ToDo will be on hold until our torch frontend's Tensor class has been updated to be able to track gradients.

from ivy.

Hi @AnnaTz thank you for your fast response,
What updates will be happen there? is there an issue for that?
I would like to see if I can contribute on that part though.

from ivy.

What updates will be happen there? is there an issue for that?

Given that gradients in Ivy are functional, there isn't really a concept of enabling or disabling them globally, so we'll need to handle this behavior in the frontend. We plan on recording the functions called if requires_grad is set True in the frontend tensor and for computing gradients we'd just redo these operations in the function that we pass to ivy.execute_with_gradients. This is similar to how we deal with views in the torch frontend. The no_grad mode would simply prevent the functions called on a tensor from being recorded, and so on for the rest of the functions in this ToDo.

This requires implementing new methods/properties in the torch frontend's tensor class, as well as writing function decorators to enable the read/write operations on the recorded gradients.

Given that this is a complex task, we would have a dedicated engineer in our team take this on, rather than making this a public task.

from ivy.